Module kernel 

Numeric kernels for identifiability-theorem diagnostics.

The kernels return scalar facts for iVAE auxiliary richness, decoder Jacobian sparsity, and manifold-SAE anchor coverage. Rust, Python, and CLI layers turn those facts into user-facing reports.

Structs

AnchorConsistencyMetrics

Scalar facts about the per-atom anchor structure of an assignment matrix.

AnchorConsistencyPreconditions

Typed pass/fail verdict for the anchor-consistency identifiability check.

AnchorConsistencyReport

AuxRichnessMetrics

Scalar facts about the auxiliary covariate / latent pair feeding an iVAE.

JacobianSparsityMetrics

Scalar facts about decoder Jacobian sparsity.

Constants

ANCHOR_DOMINANCE_DEFAULT

Default anchor-dominance threshold: the exact floating-point encoding of a strict majority.

Functions

anchor_consistency_report

Run the full anchor-consistency identifiability check and return the typed verdict. assignments is (N, K); anchor_dominance defaults to ANCHOR_DOMINANCE_DEFAULT when None.

aux_richness_metrics

Compute the iVAE auxiliary-richness numeric facts.

concat_decoder_blocks

Stack a list of per-atom decoder blocks (each shape (basis_size_k, P)) column-wise into a single Jacobian of shape (P, sum_k basis_size_k). Used by the Python diagnostics dispatcher to feed jacobian_sparsity_metrics from a ManifoldSAE.decoder_blocks payload without doing the concatenation in Python.

jacobian_sparsity_metrics

Compute mean sparsity and per-sample rank of a stack of Jacobians.